DocChat - Search for Documentation

This project explores improving documentation accessibility by implementing a lightweight semantic search system over our existing Starlight (MDX) docs. The system indexes documentation content into embeddings and enables users to ask questions in natural language (including Hindi), retrieving relevant sections based on meaning rather than exact keywords. The goal is to reduce friction in finding information and make documentation more intuitive to navigate.

1) Clone the repository and move into the docs project: - cd docs 2) Install dependencies: - npm install 3) Configure environment variables in .env: - OPENAI_API_KEY= - PINECONE_API_KEY= - PINECONE_INDEX= 4) Index documentation content into vector embeddings: - npm run rag:index 5) Start the docs application: - npm run dev 6) Open the local app in browser: - http://localhost:4321 (or the port shown in terminal) 7) Ask natural-language questions in the DocChat interface. - Example: “How do I renew my policy?” - Hindi queries are also supported for semantic retrieval.

Current scope is a local prototype for demonstration and internal validation. Known limitations: - Retrieval quality depends on chunking strategy and embedding quality - No advanced reranking layer yet for final answer optimization Future plans: - Add hybrid retrieval (keyword + semantic) - Introduce reranking to improve relevance of top results - Add feedback loop/analytics to improve answer quality over time - Deploy as an internal docs assistant with authentication and monitoring

Builders behind this 503 Day project.

Dev Adnani

Senior Solution Engineer

Passionate software engineer with a knack for designing systems, writing good code, and fueling productivity with coffee and fitness.